An employee asks a straightforward question. The answer exists somewhere across hundreds of documents, but finding it takes three searches and a message to someone who knows where to look.
For businesses exploring AI with PF Labs, this is a practical starting point: making company knowledge easier to access and use. An AI assistant becomes useful when it can retrieve the information needed to answer correctly.
Retrieval augmented generation, or RAG, connects language models with relevant external knowledge. In this guide, PF Labs explores nine real world RAG examples and the architectures behind them, showing how companies apply retrieval to customer support, internal research and everyday business workflows.
What Is RAG in AI?
RAG retrieves information from connected sources and supplies it to a language model before it generates a response.
A simple example of RAG in AI is an employee asking about parental leave. The system searches approved HR documents, retrieves relevant sections and uses them to prepare an answer. The workflow follows four stages:
Receive a question → Retrieve information → Supply context → Generate an answer
Retrieval can involve semantic search, keyword search or a combination of methods. It does not always require a vector database. Its effectiveness depends on finding useful information and generating an answer that accurately reflects it.
9 Examples of RAG Applications in Business
1. DoorDash Supports Delivery Workers
Delivery issues can involve missing orders, unclear instructions or unexpected circumstances. A fixed support script cannot anticipate every situation.
DoorDash documented a RAG implementation for Dasher support that summarizes a delivery worker’s conversation and retrieves relevant knowledge base information to help answer their question. Its system also includes an LLM Guardrail and an LLM Judge to manage response quality.
Architecture: Conversation summarization, knowledge retrieval, answer generation and quality checks.
The business lesson is practical: retrieval needs evaluation and escalation mechanisms. A relevant document alone does not guarantee that the response will resolve the problem.
2. LinkedIn Connects Historical Support Issues
Support tickets contain relationships between symptoms, products and previous resolutions. Splitting them into isolated text passages can lose those connections.
LinkedIn researchers combined RAG with a knowledge graph that preserves relationships within and between historical issues. The system retrieves relevant subgraphs to support answer generation.
Architecture: Historical issue processing, knowledge graph construction, subgraph retrieval and generated answers.
The researchers reported a 28.6% reduction in median issue resolution time in their deployed customer service setting. This example demonstrates why connected records may require more than document similarity.
3. Morgan Stanley Helps Advisors Find Internal Knowledge
Financial advisors need timely access to proprietary research and operational information. Searching extensive document libraries can interrupt client conversations and preparation.
AI @ Morgan Stanley Assistant retrieves information from the firm’s internal knowledge base. OpenAI’s case study describes how expert feedback and evaluations helped refine retrieval and response quality.
Architecture: Internal document retrieval, language model responses and ongoing evaluations.
The case study reports that over 98% of advisor teams actively used the Assistant. For businesses, this illustrates the importance of testing answers against the standards of the professionals using them.
4. GitHub Copilot Retrieves Repository Context
Developers need assistance that reflects their existing codebase, documentation and implementation patterns.
GitHub explains how Copilot retrieves relevant code and documentation, then adds those snippets to the model’s prompt. Its published implementation describes several retrieval methods, including embedding similarity and text search.
Architecture: Repository indexing, relevant snippet retrieval and context augmented generation.
This makes repository knowledge available when generating responses. The broader application extends to technical documentation assistants that must understand an organization’s terminology and existing systems.
5. Grab Assists Fraud Investigations
Investigators often repeat similar analytical queries while gathering information about suspicious activity.
Grab’s A* fraud investigation bot uses RAG to select relevant predefined queries available through its Data-Arks API platform. It executes those queries and summarizes the results through Slack.
Architecture: Relevant query retrieval, API execution and result summarization.
Here, retrieval helps identify which analytical operation to run. Connected systems provide the underlying data. Businesses adopting this approach need to evaluate query selection, execution permissions and the accuracy of the final interpretation.
6. Pinterest Finds Tables for Analytical Questions
Before generating SQL, an assistant needs to identify the tables containing relevant information.
Pinterest incorporated RAG into its Text-to-SQL workflow to help users select suitable tables. It indexes summaries of tables and historical queries, retrieves candidates and uses an LLM to narrow the selection.
Architecture: Metadata indexing, similarity search, table selection, user confirmation and SQL generation.
The important decision happens before generation: finding the right data sources. Users can validate or change the proposed tables before proceeding.
7. Vimeo Makes Video Knowledge Searchable
Training sessions and recorded presentations contain useful information that employees may struggle to locate.
Vimeo documented a video question answering system powered by RAG that retrieves information from transcripts. It stores shorter transcript passages, section summaries and descriptions of entire videos.
Architecture: Transcription, hierarchical text representations, vector retrieval and timestamp linked answers.
This supports detailed questions and broader summaries. Its documented retrieval approach uses transcript information, which means it should not be assumed to understand every diagram or visual shown in a recording.
8. Amazon Supports Product Research
Buyers often describe what they need through activities, constraints and comparisons rather than exact product names.
Amazon has documented RAG within its shopping assistant, originally called Rufus and subsequently renamed Alexa for Shopping. Its explanation describes retrieving external information to inform answers about products and trends.
Architecture: Retrieved shopping information, language models and connected shopping capabilities.
A comparable B2B application could help buyers navigate complex catalogs. Accurate prices, stock availability and specifications would still depend on reliable access to current product records.
9. Claude Projects Searches Uploaded Knowledge
Project libraries can eventually contain more information than a model can process in a single prompt.
Anthropic documents how Claude Projects activates RAG as project knowledge approaches context limits. A knowledge search tool retrieves relevant information from uploaded documents.
Architecture: Project knowledge search and selective retrieval into the model’s context.
This provides a practical example for research collections and project documentation. The official guide explains the product’s retrieval behavior but does not disclose its complete backend architecture.
RAG Project Ideas for Your Business
These RAG use cases suggest three manageable starting points:
- Policy assistant: Answer employee questions using approved internal policies.
- Documentation assistant: Help support teams locate relevant product instructions.
- Research assistant: Find evidence across a defined collection of reports.
Choose a workflow with identifiable users, maintained source material and measurable outcomes. Assess retrieval accuracy separately from answer quality. Also test permissions, document updates and situations where the assistant should acknowledge insufficient evidence.
Make Business Knowledge More Accessible with PF Labs
The strongest RAG examples address a specific information problem. Support tickets, repositories and analytical tables each require different retrieval decisions.
For your business, begin with one workflow and define what a useful answer looks like. Establish how sources will stay current and how results will be checked before expanding.
Consult PF Labs about your RAG requirements and the workflow you want to improve.